To provide a comprehensive overview of stress detection and monitoring, we searched the Google Scholar database,¹ the ACM Digital Library,² PubMed database,³ and the IEEE Xplore database.⁴

For each database, the combination of keywords “wearables,” “automated,” “stress,” “monitoring,” “detection” were used. We performed a separate search in each database by combining these keywords and grouping them with appropriate concepts. The first literature search activity was performed in August 2023 and the second literature search was performed in December 2023. Therefore, studies published after this date were not included. The publication start year was not restricted in the search. However, to consider recent research approaches, we removed all papers that appeared before 2018. For each keyword, we obtained the top 50 results that were ordered based on the rank provided by the specific library's relevance algorithm. Overall, we evaluated 400 articles – 100 from each database. These 400 articles underwent the following steps. First, duplicate articles obtained from various databases were removed. Next, we screened papers based on title and abstract based on the eligibility and exclusion criteria. After this, we filtered articles that did not meet our inclusion criteria (mentioned in Section 2.2). Finally, the remaining papers were included in the review. We next describe the inclusion criteria and the number of papers remaining at the end of each filtering step. From the first search, we selected 29 papers, and later, from the second pass, we added 10 papers from the same sources, resulting in a total of 39 papers considered for the review.

There has been a vast shift in the sensing from smartphones (Sano et al., 2015; Wang et al., 2014) to wearable sensors (Sano et al., 2018) in the last few years.

Below are the detailed exclusion criteria for this review paper:

We next describe the literature search, the screening steps, and the number of articles excluded in each step. To start off, based on the literature search, we obtained n = 400 articles (100 from each database). Among these 400 articles, there were cases where the same article was returned by different databases. Overall, we observed that there were 66 duplicate articles that we removed. At the end of this step, we had n = 334 articles. Next we removed n = 122 articles that were published before 2018, leaving us with n = 212 articles. Next, based on the exclusion criteria, we removed n = 13 articles that were not related to user studies performed on human participants, n = 24 articles that did not utilize wearables, n = 71 articles that were not on stress detection or monitoring n = 16 articles that were conducted on minors (< 18 years) or elderly (>60 years), and n = 28 articles where stress detection was performed using publicly available datasets, i.e., that did not consist of a user study. We did not consider survey papers n = 21. Thus, overall n = 224 articles were filtered based on the exclusion criteria, leaving us with n = 39 articles for this review. Figure 1 presents the consort diagram of the search and screening strategy.

We next evaluate these 39 articles in terms of the wearable technology used (Section 3) and the machine learning pipeline used in these articles (Section 4). The table summarizing all the research articles considered for this review are mentioned in Supplementary material.

Biological processes of the human body result in changes in electrical, chemical, and mechanical signals. Signals produced by these processes are known as physiological signals (Cacioppo and Tassinary, 1990). For example, the heart's pumping activity generates electrical signals corresponding to the expansion and contraction of heart muscles resulting in a heartbeat pattern (Katz, 2010). Similarly, mechanical activity in the human body corresponds to movements of body parts such as lungs during respiration, and hand or leg movement during activity (Mohanavelu et al., 2017). Some physiological change or psychological arousal in the body can result in the production of body heat, which is then released in the form of sweat—another measurable physiological signal's proxy (Gerrett et al., 2018). Our body function comprises chemical activities such as digestion, metabolism, and hormone levels, which in turn can be responsible for one's physique or behavioral changes. Specific types of physiological changes can act as a proxy for the detection and monitoring of stress (Ghaderi et al., 2015). For example, the sweating rate change without any change in physical activity can be an indicator of stress (Gerrett et al., 2018). Stress levels in individuals can also be affected by environmental factors such as ambient temperature or ambient pressure (Gordon, 2003). Various physiological signals that can measure stress levels are heart activity, brain activity, electrodermal activity, blood volume pulse, skin temperature, and muscle activity (Ghaderi et al., 2015).

Let us now explore physiological signals used for stress detection and monitoring. Various physiological signals that contribute toward stress detection are heart activity, skin conductance, skin temperature, brain activity, and activity-related signals (Ghaderi et al., 2015). ECG or PPG techniques allow capturing various heart activity-related parameters such as HR, BVP, IBI, and HRV (Shrestha et al., 2023; Montesinos et al., 2019; Han et al., 2022; Lee et al., 2022). Heart rate indicates the number of times the heart beats per minute (BPM), which reflects as immediate feedback of physical activity, emotional stress or environmental factors (Shrestha et al., 2023; Campanella et al., 2023; Moser et al., 2023; Aristizabal et al., 2021; Aqajari et al., 2023; Subash et al., 2023). Heart rate is one of the physiological signals used in stress detection and monitoring. ECG measures the heart rate by capturing electrical signals using electrodes placed near the chest or arms (Bin Heyat et al., 2022; Donati et al., 2023; Akbulut et al., 2020; Chalmers et al., 2021; Tazarv et al., 2021; Sheeraz et al., 2022; Betti et al., 2017; Xefteris et al., 2023; Lee et al., 2022; Montesinos et al., 2019; Akmandor and Jha, 2017; Ashwin et al., 2022; Smets et al., 2018b; Amalan et al., 2019), whereas PPG captures heart rate by capturing optical signals using light and measuring the blood volume from reflected light (Shrestha et al., 2023; Campanella et al., 2023; Can et al., 2020). For heart rate measurement, ECG is considered to be more accurate than PPG (Pinge et al., 2022). But ECG has motion artifacts and requires putting the electrodes near the heart that is on the chest or on arms, which can lead to discomfort in individuals during long-run studies (Beeler et al., 2018). PPG, on the other hand, can collect heart rate from body positions such as wrist, neck, or finger, making it feasible for long-run studies. Other heart activity parameters include BVP which measures changes in blood volume (Montesinos et al., 2019), IBI which refers to the time interval between consecutive heartbeats, measured in milliseconds (Han et al., 2022; Tazarv et al., 2021) and HRV which is the variation in the time intervals between heartbeats. It is a measure of the variations in the IBIs (Lee et al., 2022; Rony and Ahmed, 2020; Park et al., 2018; Mishra et al., 2020; Benchekroun et al., 2022; Szakonyi et al., 2021). BVP is collected using PPG where as HRV and IBI can be collected using ECG or PPG techniques.

Stress, which is psychological arousal, results in the generation of sweat; the rate of change of sweat generation results in a change in moisture level on the skin, thus leading to a change in skin conductance (Wilkinson et al., 1964). This change in skin conductance is measured using EDA technique and is also known as GSR. Skin conductance also contributes toward stress detection and monitoring (Anusha et al., 2019; Carreiro et al., 2020; Campanella et al., 2023; Can et al., 2020; Ashwin et al., 2022; Can et al., 2019b; Betti et al., 2017; Amalan et al., 2019). Along with skin conductance, researchers also used skin temperature (Carreiro et al., 2020; Akbulut et al., 2020; Li and Sano, 2020; Smets et al., 2018b; Shajari et al., 2023; Rachakonda et al., 2020), respiration rate (Xefteris et al., 2023; Montesinos et al., 2019; Momeni et al., 2021), activity related signals (Carreiro et al., 2020; Tazarv et al., 2021; Shrestha et al., 2023; Xefteris et al., 2023; Smets et al., 2018b; Rachakonda et al., 2020; Rodrigues et al., 2018; Han et al., 2022). Other physiological signals used for stress detection and monitoring are blood oxygen level (Akbulut et al., 2020; Akmandor and Jha, 2017), blood pressure (Akmandor and Jha, 2017), EEG—a measurement of brain activity (Sheeraz et al., 2022; Betti et al., 2017; Lee et al., 2020; Ashwin et al., 2022), and EMG—a measurement of muscle activity (Mishra et al., 2020).

Stress detection and monitoring requires capturing various body-generated signals from the individuals, as discussed in the previous subsection. Wearable devices with embedded sensors allow automatic collection of these physiological signals (Vijayan et al., 2021). Let us now look at wearable devices used for stress detection and monitoring.

Traditional medical ECG is captured using 12 leads (Khunti, 2014). This process of capturing the accurate ECG signal is sensitive toward the placement of these electrodes on specific parts of the body, such as the chest, arms, and legs. Therefore, the process of capturing ECG using the 12-leads can be tedious and time-consuming (Tomašić and Trobec, 2013). Another issue with traditional 12-lead ECG is that the setup is immobile, which makes it unsuitable for continuous monitoring of heart activity. Furthermore, traditional 12-lead ECG provides information that is usually interpreted by trained experts (Tomašić and Trobec, 2013). To avoid these issues, many healthcare companies have developed wearable devices with a varying number of leads to capture ECG and make it efficient and portable (Polar, 2024).

Chest-worn off-the-shelf devices: There are numerous chestbands available for commercial use. Along with heart activity, such chestbands are capable of collecting signals such as respiration, an inertial sensor, skin conductance, ambient temperature, and ambient pressure. For example, a Polar chestband can collect heart activity using ECG (Polar, 2024). Many researchers used earlier versions of the Polar chestband Polar H7 (Mishra et al., 2020) and recent Polar H10 (Lee et al., 2022; Zhao et al., 2023; Montesinos et al., 2019). Another chest-worn device that can capture ECG using one lead is Garmin HRM Dual (Garmin, 2024a). Similarly, Shimmer 3 ECG is another such chest-worn device that captures ECG with four leads (Shimmer, 2024). It also has 10 DoF (Degrees of Freedom) with an accelerometer, gyroscope, magnetometer, and air pressure sensor. Researchers used a similar device, Zephyr BioHarness 3 (BH3), which can capture one lead and other physiological signals such as respiration, body temperature, and body movement (Zephyr, 2024; Betti et al., 2017). A chest band from Autosense: a wireless system dedicated to stress detection captures 2-lead ECG along with respiration, ambient temperature, skin temperature, and accelerometer (Hojjatinia et al., 2021). Another chestband from BIOPAC measures EGG and respiration (BIOPAC, 2024; Kim et al., 2020). Various devices and their sensing capabilities are listed in Table 1.

Wrist-worn off-the-shelf devices: Another commonly used class of wearables are wrist-worn devices. Several wrist-worn devices are embedded with the PPG sensor (Biswas et al., 2019). PPG measures heart rate by emitting light toward the skin and measuring changes in the blood volume from reflected light (Allen et al., 2021). PPG can be measured accurately from different body parts such as the forehead, ear, nose, neck, finger, wrist, or arm (Allen et al., 2021). Unlike the chest-worn devices, the wrist-worn devices capture heart activity by using PPG technique (Biswas et al., 2019). Researchers used wrist-worn devices such as Empatica E4 (Campanella et al., 2023; Moser et al., 2023; Carreiro et al., 2020), Fitbit (fitbit, 2024), Samsung Gear Sports Smartwatch (Samsung, 2024), and Garmin Vivosmart 4 (Garmin, 2024b) for stress detection and monitoring.

Other off-the-shelf devices: Researchers experimented with other devices, such as sensorized garments i.e., smart garments, to collect ECG and EDA signals (Bin Heyat et al., 2022; Donati et al., 2023; Xefteris et al., 2023; Wu et al., 2018). Researchers also used EEG signals, which are captured by sensing near the brain. Researchers used existing devices such as MindWave Mobile EEG headset (Betti et al., 2017) for collecting EEG signals. The wrist-worn devices and their sensing ability are listed in Table 2.

The sensing capabilities of all the wrist-worn devices are almost identical except for a few sensors, such as skin temperature, skin conductance or ambient pressure. The most commonly used device, Empatica E4, allows effortless data collection and retrieval via the suite provided, where the data can be downloaded at the click of a button, unlike Garmin Vivosmart 4, which needs a Bluetooth application to receive the data from the watch. Empatica E4 is costlier as compared to other devices such as Garmin Vivosmart 4, Fitbit, or Samsung Gear Watch (Empatica, 2024; Garminvivo, 2024). Researchers need to take into account the trade-off between the cost and effort involved in data collection when choosing these devices.

The chest-worn devices capture heart rate using ECG and hence are considered as more accurate as compared to PPG (Bent et al., 2020). But chest-worn devices are sensitive to motion artifacts which can result in the collection of noisy data (Rasti et al., 2023). Although the chest-worn devices are considered to be accurate, they may not be comfortable for long-run studies (Cosoli et al., 2020). In such cases, wrist-worn devices are considered for data collection. Additionally, wrist-worn devices have built-in memory, which helps to collect and store the data within the device, whereas most off-the-shelf chest-worn devices need an application to constantly receive the data. One major challenge in using the off-the-shelf device is that these devices may not have all the necessary sensors of interest in one device. Hence, researchers develop devices with custom sensors, which we will discuss next.

Custom-made devices: Custom-made devices have a micro-controller, which can control and synchronize between the various sensors connected to it. Researchers have developed their own devices with sensors like ECG (Akbulut et al., 2020; Sheeraz et al., 2022), PPG (Zubair and Yoon, 2019), EDA (Subash et al., 2023; Shajari et al., 2023) skin temperature (Rachakonda et al., 2020; Subash et al., 2023), EEG (Sheeraz et al., 2022; Lee et al., 2020), and IMU (Rachakonda et al., 2020). However, deciding on whether to use off-the-shelf devices or develop a custom-made device is dependent upon physiological signals of interest. Using either of these devices has its advantages and challenges and can be application dependent. The off-the-shelf devices might not have all the sensors of interest or expected sampling frequency. In turn, one might have to consider using more than one wearable device or develop a custom-made device with the sensors of interest (Rachakonda et al., 2020; Zubair and Yoon, 2019; Akbulut et al., 2020).

Summary: Research studies for stress detection and monitoring use both off-the-shelf devices or custom-made devices as shown in Figure 2. Several researchers used off-the-shelf devices to collect physiological signals for stress detection and monitoring (Mishra et al., 2020; Campanella et al., 2023). In the research articles considered in our review, the off-the-shelf devices are majorly (85%) chest-worn or wrist-worn, with some researchers using other body-worn devices such as headset. Chest-worn devices placed near the heart are prominently used to observe heart activity using ECG, and Wrist-worn devices placed at various parts such as finger, wrist, and ear can collect information about heart activity, accelerometer (movement of an attached body part) and also skin conductance (Cosoli et al., 2020). This section helps to understand wearables and physiological signals used for stress detection and monitoring answering RQ1.

As described in Section 3.2, various off-the-shelf or custom-made devices can be used for data collection in user studies that aim to detect and monitor stress. Data collection also involves the collection of ground truth needed for developing the model. Data can be collected either from a controlled environment or from a free-living study. For controlled studies, researchers analyze the physiological signals either by inducing stress in a laboratory environment or by observing stressful activities that an individual performs. Orthogonally, in free-living studies, researchers provide one or more devices to participants, who wear these devices continuously while performing their normal everyday activities.

Laboratory studies: Data collection in a controlled environment is conducted by inducing stress via stressors. Researchers use a combination of these stressors with interleaved rest periods between such stressors for detecting and monitoring stress in a controlled environment. Researchers use a well-defined combination of stressors such as TSST. Several researchers used the Trier Social Stress Test (TSST; Allen et al., 2017) which has stressors such as preparing and presenting a short speech and mental arithmetic tasks with interleaved resting periods (Chalmers et al., 2021; Rodrigues et al., 2018; Rachakonda et al., 2020; Momeni et al., 2021).

In addition to the stressors introduced by TSST, researchers have also introduced additional stressors in their studies, such as watching a horror movie (Lee et al., 2022). The original TSST approach reduced residual stress by asking participants to rest between stressors. However, some researchers replaced the rest period with a short neutral video presented to the participants (Zhao et al., 2023).

Other than TSST, many researchers have used different combinations of stressors. Campanella et al. conducted a study with five different stressors or tasks with rest periods in between (Campanella et al., 2023). These tasks included the assembly of a Lego puzzle with instructions, without instructions, and a Resume presentation. Can et al. collected data from a controlled study, where the stress is induced by listening to the lectures, followed by an exam, which is similar to the mental arithmetic task in TSST (Can et al., 2020). Sheeraz et al. and Xefteris et al. used laboratory protocol with two stressors—Mental Arithmetic Task and Stroop Color-Word Test (SCWT). In the mental arithmetic task, participants were instructed to keep subtracting value 17 from 1,000 for a fixed duration of ~5 min (Mishra et al., 2020; Pinge et al., 2022), whereas in the SCWT (Scarpina and Tagini, 2017), where a word was presented to the participant, the participant was instructed to speak aloud the color of the word, and not the actual word (Sheeraz et al., 2022; Xefteris et al., 2023). For example, the word “Red” might have been written using the color purple. Participants were instructed to say aloud the word “Purple” (the color) and not “Red” (the word). Lee et al. (2020) also used a memory search stressor for inducing stress along with mental arithmetic and SCWT stressors (Sheeraz et al., 2022). Akbulut et al. (2020) collected physiological data using a stressor where participants had to watch a video with different emotions such as sadness, fear, anger, disgust, calmness, and happiness along with physical stressors – walking. Similar to watching videos containing different emotions as stressor was used by Wu et al. (2021) to collect physiological data. Betti et al. (2017), Xefteris et al. (2023), and Mishra et al. (2020) conducted lab studies where they employed two stressors – one stressor was a short presentation followed by a second stressor—a cold pressor task. Similarly, Shajari et al. (2023) used physical activity, 30–60 min of walking on the treadmill, as a stressor for collecting physiological signals. Moser et al. (2023) used acoustic stressors with 60–90 s of air horn which was amplified by speakers.

To summarize, Stressors can be broadly categorized as psychological and physical (Roldán-Rojo et al., 2021). The various psychological stressors used by various researchers are mental arithmetic task, short presentation, watching horror movie (Lee et al., 2022; Akbulut et al., 2020), Lego Puzzle (Campanella et al., 2023), Resume Presentation (Campanella et al., 2023), SCWT (Sheeraz et al., 2022; Xefteris et al., 2023), memory search task (Lee et al., 2020), acoustic stressors (Moser et al., 2023) and physical stressors include walking/running (Akbulut et al., 2020; Shajari et al., 2023), and cold pressor tasks (Mishra et al., 2020).

Free-living studies: Free-living data collection provides a naturalistic view of a person's life; the stressors here are usually not artificially provided but are natural—based on everyday activities. In the stress detection and monitoring literature in free-living conditions, the duration of such studies varies from a few hours, days or months (Smets et al., 2018a). Longer free-living studies usually assist in monitoring fine-grained activities and stressors more accurately. Data from long free-living studies help develop more accurate and personalized stress detection models.

Medical emergencies are known to induce stress in individuals. Researchers have used medical emergencies to understand free-living stress. Anusha et al. (2019) and Amalan et al. (2019) collected data for 3 h to understand pre-surgery stress (Betti et al., 2017). Bin Heyat et al. (2022) collected data for 12 h with usual activities allowed within lab. Hojjatinia et al. (2021) collected data for 3 days, Carreiro et al. (2020) collected data 4–5 days using wrist-worn Empatica E4 device. Park et al. (2018) collected physiological signals for 7 days, but data was collected three times a day for 30 s but was not collected continuously. Aqajari et al. (2023) collected data for 2 weeks. Shrestha et al. (2023) collected data for 30 consecutive days, whereas Li and Sano (2020) collected data for 30–90 days from each participant for stress detection and monitoring. There can be challenges involved while collecting data in free-living studies. These challenges involve either technical errors from wearable devices used for the data collection or human error while starting and stopping the collection on a daily basis. Longer duration of free-living studies ensure robust data collection and also help to gain useful insights from the physiological responses to various situations in day-to-day life. This also helps to identify the underlying cause of environmental conditions that can trigger stress in an individual.

Ground truth collection approaches: Data collection phase involves collection of ground truth as well. Ground Truth collection is an important task prior to the training of a model. During the laboratory study, the periods for which stressors are provided to participants are labeled as “Stressed” and all other periods are labeled as “Not Stressed.” For free-living studies, ground truth can be collected using EMA (Li and Sano, 2020; Tazarv et al., 2021). EMA is a method that involves sending questions to the users and collecting in-situ information. Researchers often use existing questionnaires such as DASS (Bin Heyat et al., 2022; Chalmers et al., 2021), STAI (Anusha et al., 2019), and PSS (Park et al., 2018) to gather necessary stress-related information. Some researchers send questions based on the information needed. For example, Li and Sano (2020) collected self-reports via email, where they collected details of mood (sad or happy), health (sick or healthy) and stress level (stressed-calm). Aqajari et al. (2023) collected self-reports by sending questions based on time or activity with seven number of EMA's per day.

The performance of the stress detection model is dependent upon the ground truth. There are challenges involved in using the collected ground truth. For laboratory studies, most of the researchers use protocol for ground truth collection. Although it is considered more accurate, sometimes it could be misleading. For example, during rest periods of laboratory study, if a user is continuously thinking/worried or, in another way, if some stressors are not capable of inducing stress, then such situations can have an impact on the performance of the stress detection model. On the other hand, the challenges with free-living studies are users might not respond accurately or not respond at all. In such cases, training the model without ground truth labels becomes challenging. To overcome such problems, researchers sometimes ignore part of the rest period data (to remove residual stress) or provide incentives for EMA responses to the participants (Mishra et al., 2020).

Sensor data collected from wearable devices may contain noise because of the presence of the body, the sensor's movement, and environmental noise. The most commonly used pre-processing steps involve noise and outlier removal, interpolation, and normalization. Physiological signals obtained from various sensors such as ECG, PPG, EDA, Skin Temperature, and Accelerometer are passed through different filters to remove noise. Researchers filtered ECG and PPG signals using Butter-worth filter (Akmandor and Jha, 2017; Benchekroun et al., 2022; Aqajari et al., 2023). ECG signal are also filtered using Fast Fourier Transform (Bin Heyat et al., 2022; Li and Sano, 2020; Akmandor and Jha, 2017), Discrete Wavelet Transform (Akbulut et al., 2020; Benchekroun et al., 2022; Moser et al., 2023), and Chebyshev II (Campanella et al., 2023). Some researchers used a bandpass filter to remove the noise and smoothed the signals using average filter (Ashwin et al., 2022; Park et al., 2018; Campanella et al., 2023). These filtered signals are then passed to the peak detector to get heart rate signals (Akbulut et al., 2020; Rodrigues et al., 2018; Campanella et al., 2023; Tazarv et al., 2021; Wu et al., 2021; Zhao et al., 2023).

EDA signal is comprised of two components: Phasic and Tonic. The phasic component is known as Skin Conductance Response(SCR). The tonic component is known as Skin Conductance Level (SCL; Posada-Quintero and Chon, 2019). For EDA, most of the researchers used the Butter-worth filter of order 4 (Anusha et al., 2019; Tazarv et al., 2021; Betti et al., 2017) for each component.

After noise removal, some researchers removed outliers (Mishra et al., 2020; Wu et al., 2021; Lee et al., 2022). The removal of noise and outliers may result in missing values for some data points. Researchers use interpolation in such cases. The missing data are interpolated using Cubic Spline (Lee et al., 2022; Mishra et al., 2020) and linear interpolation (Hojjatinia et al., 2021; Can et al., 2020; Moser et al., 2023). Some researchers normalized the data before computing features (Li and Sano, 2020; Mishra et al., 2020). Pre-processing steps are dependent upon the type of features computed. For example, most of the researchers use Fast Fourier Transform to convert the data from time-domain to frequency domain (Moser et al., 2023; Amalan et al., 2019; Aqajari et al., 2023; Akbulut et al., 2020).

The next step in the process of stress detection and monitoring is feature computation. We will now look at the different types of features computed using various physiological signals.

Researchers compute time-domain and frequency-domain features from physiological signals such as ECG, PPG, EDA, Skin Temperature, and Accelerometer. Various time-domain and frequency-domain features are computed from the sensor data obtained from various wearables. Computing frequency domain features involves transforming time series data into the frequency spectrum using mathematical transformations like the Fourier Transform (Ashwin et al., 2022). These features help in understanding how the signals are distributed across different frequency bands.

As discussed in the earlier section, ECG and PPG can give physiological signals such as heart rate (HR), R-R intervals (RR), also known as heart rate variability (HRV), and inter-beat interval. Researchers compute statistical time-domain features such as mean of HR or RR intervals, standard deviation of HR or RR intervals, minimum, maximum, median of heart rate or RR intervals (Akbulut et al., 2020; Mishra et al., 2020; Rodrigues et al., 2018). Other time-domain features include RMSSD (Root mean square of successive differences between the adjacent RRs), SDSD (Related standard deviation of successive RR interval difference; Amalan et al., 2019; Aqajari et al., 2023; Benchekroun et al., 2022; Ashwin et al., 2022; Lee et al., 2022; Rony and Ahmed, 2020), NN50 (Number of successive RR intervals pairs that differ more than 50s), PNN50 (NN50 divided by the number of RR intervals), SDNN (Standard Deviation of normal RR intervals; Amalan et al., 2019; Montesinos et al., 2019; Tazarv et al., 2021; Betti et al., 2017), HRV triangular index (total number of RR intervals divided by the height of the histogram of all RR intervals measured on a scale with bins of 1/128 s), TINN (triangular interpolation of RR interval histogram; Can et al., 2019b; Akbulut et al., 2020). Researchers also computed breathing rate using PPG signals (Aqajari et al., 2023; Tazarv et al., 2021).

Researchers computed frequency-domain features by using band pass filter for VLF (very low frequency; Chalmers et al., 2021; Betti et al., 2017; Momeni et al., 2021), LF (low frequency; Benchekroun et al., 2022; Can et al., 2019b; Wu et al., 2021), HF (high frequency) and LF/HF Ratio (Chalmers et al., 2021; Wu et al., 2021; Betti et al., 2017; Wu et al., 2018). Some researchers computed total power (Benchekroun et al., 2022; Amalan et al., 2019), energy (Can et al., 2020) pLF (prevalent low-frequency oscillation of heart rate), pHF (prevalent high-frequency oscillation of heart rate; Can et al., 2019b, 2020).

Researchers compute statistical, phasic, and tonic features of EDA. Statistical features of EDA include mean, median, mode standard deviation, maximum, and a minimum of EDA (Campanella et al., 2023). Phasic features capture rapid transient changes as a response to the situation, whereas tonic features vary slowly over time, which helps to understand psychological state over a longer period of time (Cecchi et al., 2020). Features of phasic and tonic components are mean and standard deviation of Phasic component (Wu et al., 2021; Betti et al., 2017), peaks phasic (Can et al., 2020) mean and standard deviation of tonic component (Can et al., 2020), Skin conductance response is a rapid, transient change in skin conductance in response to the events. Hence, researchers computed features such as startle (number of detected startles), Startle mean (mean of the amplitude of the startles), Startle SD (standard deviation of the amplitude of the startles), Rise time mean (mean of the rise time of the startles), Rise time SD (standard deviation of the rise time of the startle), Fall time mean (mean of the fall time of the startles), and Fall time SD (standard deviation of the fall time of the startles) (Wu et al., 2021; Betti et al., 2017).

Some researchers computed features such as mean of respiration duration, SD of respiration signal, and median of respiration duration for respiration signal (Akmandor and Jha, 2017; Momeni et al., 2021). Other features computed are mean blood pressure, mean blood oximeter (Akmandor and Jha, 2017), mean X, mean Y, mean Z, and magnitude for accelerometer signals (Can et al., 2019b). Overview of the features computed are shown in Table 3.

Researchers perform stress detection and monitoring either by developing classification models or by performing statistical analysis. The features computed from various physiological signals extract meaningful information. from the raw signals. These features are then passed on to train the machine learning model. Instead of developing a classification model, few researchers analyze the data collected using wearables using statistical analysis (Rony and Ahmed, 2020). Some of the commonly used statistical tests used are Paired-test, Anova, or Pearson's Correlation (Chalmers et al., 2021; Amalan et al., 2019).

Daily life situations or work pressure can cause chronic stress or frequent episodes of acute stress. In order to cope with such stress, individuals often get involved in substance use disorder or other harmful habits such as smoking, excessive alcoholism, or even overeating. Detecting and monitoring stress helps in keeping the wellbeing of individuals. Researchers considered stress in different areas such as daily habits (Tazarv et al., 2021; Han et al., 2022; Moser et al., 2023), stress experienced at the workplace (Donati et al., 2023; Xefteris et al., 2023), in college students (Li and Sano, 2020; Park et al., 2018; Can et al., 2019b), or while driving (Rony and Ahmed, 2020). Some researchers detected stress in patients before undergoing surgery (Anusha et al., 2019; Amalan et al., 2019), patients with metabolic syndrome (Akbulut et al., 2020) or cravings during treatment for substance use disorder (Carreiro et al., 2020; Shrestha et al., 2023) and stress levels before and after smoking events (Hojjatinia et al., 2021; Akbulut et al., 2020).